Advancing Artificial Intelligence with Neural-Symbolic Learning and its Application to Generative Policies in Distributed Coalition Operations
Military / Coalition Issue
Coalition operations may require rapid reconfiguration for a given task. Devices, systems and services (hardware and software) owned by different coalition partners will be required to integrate and operate seamlessly as a unified system. Furthermore, there is a lack of communication to back-end compute infrastructure, malicious adversaries present, and there is a requirement for human-machine interaction to support troops and deployed personnel. The coalition environment creates an additional challenge of trust, where parties within the coalition may not fully trust one another.
Core Idea and Key Achievements
This demonstration presents a policy learning framework which enables devices to be rapidly brought online into a mission environment. Devices can autonomously learn their behaviour, called policies, through local communication to nearby devices. Policies can be learned from a small number of examples and are explainable to humans. The demonstration includes the Answer Set Grammar Generative Policy Model which enables autonomous behaviour through the generation of policies in varying operational contexts, where the generative model is learned from policy examples. The demonstration also includes the Neural-Symbolic Generative Policy Model, which extends the framework to learn from unstructured data present in a military environment (e.g. imagery, video, audio, sensor data). The Neural-Symbolic extension is robust to distributional shift in input data during learning, reducing the likelihood of an adversarial attack.
Specifically, we have integrated FastLAS into a neural-symbolic architecture, called NSL, that extracts symbolic features from heterogeneous contextual data, by means of pre-trained Deep Learning (DL) systems, and uses these features to learn context-aware optimal patterns of behaviour. We have demonstrated that the learned models are highly accurate, even when the DL predictions are highly uncertain, are interpretable, require 3-4 orders of magnitude less training data, and are 3-4 times faster than pure DL approaches.
Implications for Defence
Both FastLAS and NSL systems are loosely coupled (they can integrate multiple Artificial Intelligence (AI) systems for feature extraction from unstructured data, and different symbolic learning solvers), and require limited computational power. As a result, these systems can perform ‘edge of network’ learning for a range of tactical applications (e.g., network management, logistics, sensing).
Enables Defence to learn and adapt, in near-real time, patterns of behaviour from streaming data (including network, ISTAR (intelligence, surveillance, target acquisition, and reconnaissance) and HUMS (health and usage monitoring systems) data). Enables Defence to outperform enemy forces by adopting AI systems capable of autonomously learning new AI services that efficiently process information quicker and more intelligently at the ‘pace of the fight’.
Readiness and Alternative Defence Uses
This work is a technology readiness level (TRL) 4. FastLAS and NSL are available as open source software. While the approach has been demonstrated on images and structured data, it is applicable in general to other time-series data (e.g. network traffic, access control policies, decision-making strategic policies).
Resources and References
- Law, Mark, et al. “Representing and learning grammars in answer set programming.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 33. No. 01. 2019.
- Law, Mark, et al. “FastLAS: scalable inductive logic programming incorporating domain-specific optimisation criteria.” Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 34. No. 03. 2020.
- Cunnington, Daniel, Alessandra Russo, Mark Law, Jorge Lobo, and Lance Kaplan. “NSL: Hybrid Interpretable Learning From Noisy Raw Data.” arXiv preprint arXiv:2012.05023 (2020).
Imperial College London, IBM UK, Purdue University, ARL.